🌌 Digital Habits & Mental Health — Wellbeing Risk

Production-minded notebook for binary wellbeing risk (high_risk_flag ∈ {0,1}) derived from digital behavior signals.

---

🎯 Goal

Predict a calibrated wellbeing risk score in [0,1] from daily digital patterns — screen time, unlocks, sleep, notifications, activity.

Focus:

• Stable CV, not leaderboard luck
• Minimal, meaningful visuals
• Zero leakage · Deployable artifacts

---

🧠 End-to-End Pipeline

1. Data Loading → clean read, column sanitization, dedup, schema snapshot
2. EDA (Focused) → leakage guard · top Spearman correlations · compact Plotly visuals
3. Feature Engineering → ratios & interactions · winsorized z-scores · log1p transforms
4. Models → Logistic Regression · Random Forest · XGBoost (auto GPU/CPU)
5. Calibration → isotonic reliability fit on the best model
6. Decisioning → min-cost threshold + capacity constraint (optional review rate)
7. Interpretability → AUC-based permutation importance (Kaggle-safe)
8. Export → calibrated model + schema + metadata + inference smoke test

---

🔬 Features & Cross-Validation

Key features (examples):

• screen_per_sleep_norm = device_hours_per_day / sleep_hours
• unlocks_per_hour = phone_unlocks / (device_hours_per_day + 0.1)
• social_ratio, work_ratio from time splits
• Winsorized z-scores for device_hours, sleep_hours, stress, anxiety, depression, happiness
• log1p transforms for unlocks, social_media_minutes, work_minutes, notifications_per_day

---

📈 Metrics & Visualization

• ROC–AUC · Average Precision · Brier Score
• Comparative ROC/PR curves (all models)
• Reliability curve (calibrated best) · Lift curve
• AUC-based permutation importance for interpretability

---

🧮 Cost-Aware Decisioning

Threshold grid t ∈ [0,1] using:

total_cost = FP*C_fp + FN*C_fn - TP*benefit

• Report min-cost and best-F1 thresholds
• Optional capacity constraint → choose lowest-cost threshold under review limit

---

📂 Outputs

All reproducible artifacts stored under artifacts/:

eda/        → schema_overview.csv, engineered_corr.csv
metrics/    → leaderboard_metrics.csv, permutation_importance_auc.csv
decision/   → decision_cost_curve.csv, thresholds_summary.csv
models/     → best_calibrated.joblib, feature_schema.json, model_info.json
inference/  → smoke_test_report.json

---

⚡ Quick Start

pip install -r requirements.txt

Plotly renderer:

import plotly.io as pio
pio.renderers.default = "notebook_connected"  # auto PNG fallback (kaleido)

---

🗂️ Dataset

• Target: high_risk_flag (0/1)
  If categorical → TARGET_MAP = {"Low":0, "High":1}
• Default path: /kaggle/input/digital-health-and-mental-wellness/Data.csv

---

📁 Repo layout

.
├── predicting-wellbeing-risk.ipynb
├── data/
│   └── raw/               # put Data.csv here for local runs
├── artifacts/             # exported tables / metrics / model files
├── repo_utils/
│   └── pathing.py         # local + Kaggle path helpers
├── CASE_STUDY.md
├── requirements.txt
└── .gitignore

---

🧭 Data loading (local + Kaggle)

Local (recommended):

• Put Data.csv under data/raw/

Kaggle:

• Falls back to /kaggle/input/digital-health-and-mental-wellness/Data.csv

Optional override:

• Set DATA_PATH to a full file path (local runs).

---

🧾 Case Study

See CASE_STUDY.md for the project story, decisions, and takeaways.